Bump validated SGLang stack to 0.5.20

[BaoYunkai]

Summary

• Raise Hyperloom's validated SGLang defaults to 0.5.20 (SGLANG_REF, SGLANG_PRETEND_VERSION, recommended lmsysorg/sglang-rocm:v0.5.20-rocm724-* tags).
• Refresh compatibility/install docs and examples; document patchless kernel-shape profiling via TraceLens kernel_shape_tool.

Test plan

• pytest src/hyperloom/inference_optimizer/tests/test_baremetal_doc_version_consistency.py

[xiaofei-zheng]

Blocking

1. The merge commit 415afca0 discards everything on main since the branch point — merging this PR reverts 6 merged PRs.

415afca03 ("Merge remote-tracking branch 'origin/main'") has a tree identical to its first parent:

$ git diff --stat 8949dd2e4 415afca03
(empty)
$ git diff --name-status d658bb480 415afca03 | wc -l
60

None of main's side of the merge survived. Everything merged between the branch point 6b4b4137 and d658bb48 would be rolled back by merging this PR: #1509 (ATOM detailed profiling), #1591 (preserve operator GEAK checkouts), #1589 (retire CHANGELOG.md), #1549 (ROCm 10 e2e), #1555 (single definition of agent default model ids), #1515 (catalog entry skill).

Concrete examples in the PR diff:

• src/hyperloom/common/llm_config.py — DEFAULT_CLAUDE_MODEL / DEFAULT_CODEX_MODEL and the backend-decides-the-last-rung resolution from #1555 are removed, and resolve_forge_llm_model goes back to a caller-supplied default="", so an OAuth-token run with CLAUDE_MODEL unset posts an empty model id again.
• CHANGELOG.md comes back (+2286) and docs/release-notes.md loses 669 lines, reverting #1589.
• .github/workflows/federated-skill.yml, scripts/tests/test_federated_skill_contract.py, examples/skills/hyperloom-workload-optimizer/SKILL.md, src/kernelforge/tests/test_rocpc_profile_discovery.py are deleted.
• src/hyperloom/inference_optimizer/assets/install.sh (-160), rocpc_profile.py (-141), _workload_envs.py (-81) are rewound to their pre-main versions.

This also silently drops the reason main currently pins 0.5.19: main's install_baremetal.sh/compatibility.rst state the pin sits at 0.5.19 because through 0.5.18 the HIP extra pinned compressed-tensors==0.15.0, capping torch below 2.11 and making ROCm 10 unresolvable. The 0.5.20 pin here is fine on that axis, but that constraint should be re-stated on top of current main, not erased.

CI is green because the branch tree is internally consistent with the old main — it does not catch the revert.

Please redo the integration (git merge origin/main from a clean branch, or rebase the three real commits onto current main) so the diff contains only the SGLang 0.5.18 → 0.5.20 change on top of main's current 0.5.19 state.

2. No docs/release-notes.md entry.

The validated default framework stack (SGLANG_REF, SGLANG_PRETEND_VERSION, recommended lmsysorg/sglang-rocm image tags, models.tsv, quick-start Dockerfile) is user-visible, and docs/release-notes.md "Unreleased" is where main carries such changes (#1580 added one for a comparable change). Add an entry, including the upgrade note that users on the v0.5.18-rocm724-* images need to move to v0.5.20-rocm724-*.

Checked

• Real PR commits (5ee35eec, 8949dd2e, 6751302) reviewed line by line: docs/examples tag updates, install_baremetal.sh pin + IMAGE_HINT, models.tsv, quick-start Dockerfile, and the relaxed test_baremetal_doc_version_consistency.py. Doc/asset tags are consistent across all nine touched files (mi30x-20260919 / mi35x-20260920).
• SGLANG_REF=d158602ff1d2cb953196c95158c488d503d2470c resolves to the annotated tag object v0.5.20 (commit 94602c9c2b), so SETUPTOOLS_SCM_PRETEND_VERSION_FOR_SGLANG=0.5.20 matches. Worth confirming git fetch --depth 1 origin <tag-object-sha> in install_sglang_from_source behaves as expected, since the surrounding comment assumes a commit SHA.
• Shape-discovery gate: _server_patcher.py routes >= 0.5.18 to the no-patch kernel_shape_tool, so dropping the "patches fail on the tag" rationale is correct for 0.5.20.

[BaoYunkai]

Superseded: branch was rebuilt on current main (removed the bad merge that dropped main-side changes). A new PR follows.